Pinned post
Latest posts 5
I had to split the body and head for 3d printing because the head is massive. First printing test with half size looks cool! My kids love it and they want their own custom funko pops.
Tried to create my own Funko Pop with Astra today. I first generated reference views in ChatGPT, then gave Codex access to Blender. It's not bad, but we are not there yet. This is a verifiable domain, so I would expect fast progress in the coming months.View quoted post on X
Tried to create my own Funko Pop with Astra today. I first generated reference views in ChatGPT, then gave Codex access to Blender. It's not bad, but we are not there yet. This is a verifiable domain, so I would expect fast progress in the coming months.
I wonder how much of the gains are due to algorithmic efficiency versus training a model specialized in software engineering and comparing only on software engineering data. A specialized model will always use fewer resources than a general one.
Frontier pretraining is said to be a big-lab-only game. We don’t have 100k chips yet, so there’s only one way: algorithmic efficiency. Our new recipe matches DeepSeek V4 Pro’s pretrain using 50x less compute – that’s roughly half the FLOPs used for GPT3, or ~$0.5M on GB200.View quoted post on X
magic.dev/blog/pretraini…
I have created a small website to visualize temperature changes over the last 50 years for any location. ironbar.github.io/temperature-tr…
For example, in Pamplona, the number of days with a maximum temperature above 30 °C has increased from ~5 in 1977 to ~28 in 2024
For example, in Pamplona, the number of days with a maximum temperature above 30 °C has increased from ~5 in 1977 to ~28 in 2024